Blackwell Ultra: The Ultimate AI Computing Platform
NVIDIA HGX™ B300 is built on the revolutionary Blackwell Ultra architecture, featuring 288GB HBM3e memory per GPU and delivering 11x faster inference performance than H100. With 120 petaFLOPS of FP4 dense compute and 5th-generation NVLink technology, B300 sets new standards for enterprise AI training, inference, and large language model deployment.
Key Highlights
Experience breakthrough performance with Blackwell Ultra architecture, featuring 2.3TB total GPU memory, 11x faster inference than H100, and industry-leading connectivity with 5th-generation NVLink technology.
Blackwell Ultra Architecture
Built on NVIDIA's most advanced Blackwell Ultra architecture with 208 billion transistors, delivering unprecedented AI compute performance with 5th-generation Tensor Cores supporting FP4, FP6, and FP8 precision formats.
11X Faster Inference
Delivers up to 11x higher inference performance compared to H100 for large language models like Llama 3.1 405B, with 2x faster attention performance than B200.
120 PetaFLOPS Compute
Achieves 15 petaFLOPS FP4 dense compute per GPU (120 petaFLOPS total for 8 GPUs), with 30 petaFLOPS FP4 sparse compute per GPU, representing 1.5x performance improvement over B200.
5th-Gen NVLink
Features 5th-generation NVLink with 14.4 TB/s total GPU-to-GPU bandwidth and 1.8 TB/s per GPU via NVSwitch, plus 1.6 TB/s networking bandwidth (double the B200).
Technical Specifications
Complete technical specifications for the NVIDIA HGX B300 platform, featuring Blackwell Ultra architecture with 208 billion transistors, 288GB HBM3e per GPU, and 5th-generation NVLink interconnect technology.
| Technical Detail | Specification |
|---|---|
| GPU | 8x NVIDIA Blackwell Ultra B300 GPUs (208 billion transistors, TSMC 4NP process) |
| GPU Memory | 2,304GB total GPU memory (288GB HBM3e per GPU) |
| Compute Performance | 120 petaFLOPS FP4 dense (15 petaFLOPS per GPU), 240 petaFLOPS FP4 sparse (30 petaFLOPS per GPU) |
| Training Performance | 4x faster than H100 for large language models |
| Inference Performance | 11x faster than H100 for models like Llama 3.1 405B |
| Attention Performance | 2x faster than B200 |
| Power Consumption | Up to 1,100W TDP per GPU (~16-17kW system max) |
| GPU Interconnect | 5th-generation NVLink: 14.4 TB/s total bandwidth, 1.8 TB/s per GPU via NVSwitch |
| Memory Bandwidth | 8 TB/s per GPU HBM3e bandwidth |
| CPU | 2 Intel® Xeon® Platinum processors (compatible with latest generation) |
| System Memory | Up to 6TB DDR5 system memory |
| Networking | 1.6 TB/s total networking bandwidth (double B200), 4x OSFP ports, NVIDIA ConnectX-8 VPI, BlueField-4 DPU support |
| Management Network | 10Gb/s onboard NIC with RJ45, 200Gb/s dual-port ethernet NIC, Host BMC with RJ45 |
| Storage | OS: 2x 3.84TB NVMe M.2, Internal storage: 8x 7.68TB NVMe U.2 |
| Security Features | Secure AI with confidential computing, TEE-I/O capability, inline NVLink protection |
| Software | NVIDIA AI Enterprise: Optimized AI Software, NVIDIA Base Command™: Orchestration and Cluster Management, HGX OS / Ubuntu |
| Rack Units (RU) | 10 RU |
| System Dimensions | Height: 17.5in (444mm) Width: 19.0in (482.2mm) Length: 35.3in (897.1mm) |
| Operating Temperature | 5–30°C (41–86°F) |
| Enterprise Support | Three-year Enterprise Business-Standard Support for hardware and software, 24/7 Enterprise Support portal access, Live agent support during local business hours |
Ready to Transform Your AI Infrastructure?
Unlock the full potential of Blackwell Ultra architecture with NVIDIA HGX B300. With 11x faster inference, 2.3TB GPU memory, and 120 petaFLOPS compute power, B300 is engineered for the most demanding AI workloads. Contact our experts to discover how this revolutionary platform can transform your enterprise AI infrastructure.